Will the highest score achieved by a Moonshot Kimi model on Humanity’s Last Exam in 2026 be 50% or higher?

- [2026-10-07] The official Scale AI "Humanity's Last Exam" leaderboard (the market's resolution source, mirrored at labs.scale.com/leaderboard/humanitys_last_exam) shows only one Moonshot entry, "Kimi K2.5," at 24.4% accuracy — far below 50% — while the leaderboard's current leader is GPT-6 Astra at 54.8%. https://benchlm.ai/benchmarks/scale-humanitys-last-exam - [2026-10-07] On that same official snapshot the top rows are GPT-6 Astra (54.8%), Fable 5.1 xhigh (46.5%) and gemini-3.1-pro-preview thinking-high (46.4%); no Kimi model is close to the top. https://benchlm.ai/benchmarks/scale-humanitys-last-exam - [2026-09-17] The Center for AI Safety/Scale's official HLE page confirms the benchmark was refreshed via "HLE-Rolling," a continually updated fork with harder questions replacing easier ones, which tends to push scores down rather than up for any given model. https://scale.com/leaderboard/humanitys_last_exam - [2026-09-22] The official agi.safe.ai site's most recent "Latest News" item (as of today) is the release of "HLE-Diamond" on 09/22/2026, indicating the benchmark itself is still evolving, which could affect how/whether new Kimi scores are posted before year-end. https://agi.safe.ai/ - [2026-07-27] Moonshot's newest flagship, Kimi K3 (2.8T parameters), reports on its own model card an HLE-Full score of 43.5% without tools and 56.0% with tools — but this is a vendor-reported figure, not yet reflected on the official Scale/HLE leaderboard. https://devoriales.com/kimi-k3-open-weights-what-moonshot-actually-shipped - [2026-07-16] Moonshot's K3 model card (self-reported) lists HLE-Full at "43.5 / 56.0" (no-tools/with-tools), trailing rival vendor-reported scores such as Claude Fable 5 (53.3/63.0) and GPT-5.6 Sol (44.5/58.0). https://huggingface.co/moonshotai/Kimi-K3 - [2026-04-xx] Earlier in 2026, Moonshot's Kimi K2.6 vendor blog claimed "Humanity's Last Exam with tools at 54.0" — a self-reported, tools-enabled figure that, like other Moonshot vendor claims, has not appeared on the official Scale leaderboard. https://dev.to/dishant0406/kimi-k26-dropped-the-internet-went-crazy-then-reality-hit-4dgf - [2026-02-xx] A documented pattern exists of large gaps between Moonshot's self-reported HLE numbers and independently/officially measured ones: Kimi K2's vendor-claimed ~50% HLE score was independently found to be only about 29.4% under standard, non-optimized conditions. https://www.mindstudio.ai/blog/humanities-last-exam-benchmark-score-inflation-explained - [2026-11-09 (reported)] An earlier generation reasoning model, Kimi K2 Thinking, was noted by outside trackers as sitting at about 23.9% on HLE text-only (no tools) around that date, well under the 44.9%/51% "heavy" tool-assisted figures Moonshot itself publicized. https://thezvi.substack.com/p/kimi-k2-thinking - [2026-10-11] As of today, no Kimi model (K2, K2 Thinking, K2.5, K2.6, or K3) appears on the official Scale/HLE leaderboard with a score at or above 50%; the only listed Kimi row (K2.5) sits at 24.4%. https://benchlm.ai/benchmarks/scale-humanitys-last-exam BOTTOM LINE: On the actual resolution source (the official Scale/CAIS HLE leaderboard), the only listed Kimi model currently sits at 24.4% — nowhere near 50% — even though Moonshot's own marketing blogs for K2.5/K2.6/K3 claim tool-assisted scores of 50-56%; those vendor numbers have not been verified or posted to the official leaderboard, and independent re-tests of past Kimi claims have landed far below Moonshot's self-reported figures, making an official ≥50% Kimi HLE score by year-end look currently unsupported by the primary data source. NOT FOUND: No evidence that Moonshot has formally submitted Kimi K2.6 or Kimi K3 results to the official Scale/CAIS HLE leaderboard (only "Kimi K2.5" appears as of the October 7, 2026 snapshot); no official leaderboard entry corroborating any Moonshot vendor-blog claim of ≥50% HLE accuracy.

- [2026-10-11] The question's own resolution source (agi.safe.ai / Scale's official HLE leaderboard) requires sign-in and isn't publicly browsable directly; third-party mirrors of that same "Scale score" snapshot are the best available proxy. https://agi.safe.ai/dashboard - [2026-09-30] A mirror of the Scale Labs official HLE leaderboard puts Moonshot's newest and largest model, Kimi K3, at only "GPT-6 Astra OpenAI 60.6%... 10 Kimi K3 Moonshot AI · Closed 22.2%" — far below the 50% strike and well behind the leading proprietary models. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-09-23] An earlier Scale snapshot (one month prior) shows Kimi K2.5 at only "Kimi K2.5 Moonshot AI 24.4%" , confirming Moonshot models have persistently scored in the low-to-mid 20s on the Scale/official-style leaderboard even as other labs' models climbed well past 50%. https://benchlm.ai/benchmarks/scale-humanitys-last-exam - [2025-09-05 release / later inflation exposed] For Kimi K2 (2025), Moonshot's own benchmark card listed a score of roughly 50% on the Humanities Last Exam , but independent testing found 29.4% — a documented ~21-point gap between Moonshot's self-reported HLE numbers and independently verified results, which is the key base-rate precedent for discounting Moonshot's self-reported figures. https://www.mindstudio.ai/blog/humanities-last-exam-benchmark-score-inflation-explained - [2025-11-06] Kimi K2 Thinking's own release materials claimed "K2 Thinking achieves 44.9% on HLE with tools... establishing new records in multi‑domain expert‑level reasoning performance" , and only under a non-standard, heavily compute-amplified "heavy mode" (multiple parallel samples) did Moonshot report "HLE Heavy Mode: 51.0%" — a configuration unlikely to match how the official leaderboard scores single-run accuracy. https://moonshotai.github.io/Kimi-K2/thinking.html ; https://docs.api.nvidia.com/nim/reference/moonshotai-kimi-k2-thinking - [2026-10-09] Independent aggregator Artificial Analysis-based trackers (a different methodology than Scale's) place Kimi K3 somewhat higher but still below the strike: "Kimi K3: 43.5%... behind Fable 5 and Opus 4.8" on HLE without tools, versus 56.0% with tools per the same source, illustrating how much "HLE score" varies by tool-use configuration and tracker — none of which is confirmed to be the configuration the Polymarket resolution source (official agi.safe.ai leaderboard) will display. https://emergent.sh/learn/kimi-k3-benchmark - [2026-10-08] Across several leaderboard aggregators (llm-stats, modelgrep, benchlm) the top overall HLE scores in October 2026 are held by Anthropic (Claude Fable 5.1, ~59-65%) and OpenAI (GPT-6 Astra, ~54-61%), not by any Moonshot Kimi model, with Kimi K3 typically ranked around 10th-12th place among tracked models at scores in the 22-43% range depending on source. https://benchlm.ai/benchmarks/hle ; https://llm-stats.com/benchmarks/humanity's-last-exam - [Historical pattern] Since HLE's public release in January 2025, the overall leaderboard has shown rapid upward drift — top scores rose from roughly 20-25% in early 2025 to 37.5% (Gemini 3 Pro, Feb 2026) to past 50-60% by September-October 2026 for leading closed models (GPT-6 Astra, Claude Fable/Opus 5.x) — but Moonshot's Kimi line has lagged this frontier curve on independently-tracked numbers the entire time. https://letsdatascience.com/blog/humanitys-last-exam-the-test-thats-humbling-the-worlds-smartest-ai - [No market-specific odds found] No betting market or forecaster-specific history for "Kimi crossing 50% on HLE" could be located beyond a resolved Manifold market on a related but distinct question ("K2 Thinking gets top score on HLE, independent eval") which "Resolved NO" , reinforcing that independent/official evaluations have consistently rated Kimi models below the top of the field. https://manifold.markets/Bayesian/k2thinking-gets-top-score-on-hle-ac BOTTOM LINE: The base-rate evidence strongly favors "No": across the entire ~21-month history of HLE, no Moonshot Kimi model has ever posted a verified/official ≥50% score (self-reported Kimi K2 and K2-Thinking claims near or just above 50% were either independently debunked or relied on non-standard "heavy mode" aggregation), and as of October 2026 the official-leaderboard-style scores for the latest Kimi models (K2.5, K3) sit in the 22-44% range, well behind frontier leaders already at 50-65%. NOT FOUND: Could not directly view the live official agi.safe.ai/Scale leaderboard (it requires sign-in), so had to rely on third-party mirrors; could not find any explicit prediction-market or forecaster probability estimate specifically for "Kimi model ≥50% on HLE in 2026" outside of the Polymarket contract itself (excluded per instructions).

YES means: the official Humanity's Last Exam leaderboard (agi.safe.ai / Scale, or an accepted official-results substitute) shows some Moonshot Kimi model's "HLE Accuracy" figure at ≥50% at any point before December 31, 2026, 11:59 PM ET. - [2026-07-16] Moonshot's own Kimi K3 model card reports HLE-Full at 43.5 without tools and 56.0 with tool augmentation , and a third-party mirror (unsloth) separately lists a Hugging Face "cais/hle" evaluation-results badge of 56 for K3, i.e. a non-Moonshot-hosted page still crediting the model with a 56% figure. https://huggingface.co/unsloth/Kimi-K3 - [2026-04-xx] Moonshot's Kimi K2.6 model card shows a clear, steadily rising tool-assisted HLE trend across Moonshot's own last three flagships: HLE-Full (w/ tools): Kimi K2.6 54.0, versus Kimi K2.5 50.2 , both already above the 50% strike on the vendor's own tool-enabled configuration. https://huggingface.co/moonshotai/Kimi-K2.6 - [2026-04-xx] On that same K2.6 card, Moonshot's tool-augmented score also already exceeded contemporaneous frontier closed models: GPT-5.4 (xhigh) 52.1, Claude Opus 4.6 (max effort) 53.0, Gemini 3.1 Pro (thinking high) 51.4 , all clustered in the low-50s alongside Kimi, showing 50%+ with-tools HLE scores are now a routine frontier-model outcome, not a Kimi-specific anomaly. https://huggingface.co/moonshotai/Kimi-K2.6 - [2026-06-12] Moonshot's Kimi K2.7 Code release, a specialized coding-focused model, still posted a vendor-reported HLE-with-tools score of 54.0, explicitly "HLE with tools (leads GPT-5.4 and Opus 4.6) 54.0" , indicating Moonshot's tool-augmented HLE performance did not regress even in a model optimized for coding rather than reasoning. https://www.okou.ai/en/models/kimi-k2-7-code - [2026-10-09] An independent third-party aggregator (Vellum) lists Kimi K3 as the open-source "Best Overall" model on Humanity's Last Exam with a score of 56%, an aggregator-reported figure distinct from Moonshot's own blog. https://www.vellum.ai/open-llm-leaderboard - [2026-08-xx] A head-to-head comparison piece using Artificial Analysis-style data put Kimi K3's HLE-with-tools score even higher than Moonshot's own card, at "HLE w/ Tools (Hard Tasks with Tool Access): GLM-5.3 at 62.5%, Kimi K3 at 59.8%" — a 4.5% edge for GLM-5.3, meaning an independent-style tracker already shows Kimi K3 near 60% on a tool-enabled HLE configuration. https://medium.com/data-science-in-your-pocket/glm-5-3-vs-kimi-k3-a0e8694aa47d - [2026-07-23] Artificial Analysis's own tracked (non-vendor) number for Kimi K3 (Max), as mirrored by OpenRouter, puts HLE at 46.9%, already closing most of the historical gap between Moonshot's self-reported and independently measured no-tools scores (compare the ~21-point K2 gap documented earlier), making a further small jump to ≥50% on a no-tools configuration plausible with one more model generation. https://openrouter.ai/moonshotai/kimi-k3 - [2026-10-11] Moonshot's release cadence has been rapid and continuous through 2026 — K2.5 (January 2026), K2.6 (April 2026), K2.7 Code (June 2026) — followed by K3 in July, implying at least one more flagship or point release is plausible before the December 31, 2026 deadline, giving Moonshot additional opportunities to post a higher score. - [2026-10-11] The resolution criteria explicitly allow for "a clear equivalent metric if the data's presentation or terminology is restructured," which gives room for a tool-augmented or harness-specific Kimi score (already 54-60% per multiple vendor and tracker sources) to count if the official leaderboard's displayed "HLE Accuracy" column adopts or is read as a tools-enabled configuration, as several leading closed models' rank-1 scores (e.g., the current 54.8% leader) may themselves reflect a tool-assisted harness rather than a strict no-tools run. - [2026-09-22] The benchmark itself is in flux with the "HLE-Diamond" refresh; new/harder question sets historically produce a period where only the very top models clear 50%, but Moonshot has shown it can match frontier labs' tool-augmented numbers within the same release cycle (K2.6/K2.7 matching or beating GPT-5.4/Opus 4.6's low-50s tool scores), so a Kimi model landing in the mid-50s on a refreshed official board is not without precedent. BOTTOM LINE: The strongest YES path is that Moonshot's own and several third-party (Artificial Analysis/OpenRouter, Vellum) trackers already show Kimi models at 54-60% HLE under tool-augmented or specialized configurations, with a clear generation-over-generation upward trend (K2.5→K2.6→K2.7→K3) and a continuing release cadence — if either the official leaderboard adopts/accepts a tools-enabled score or Moonshot ships one more model before year-end matching its own recent with-tools trajectory, a ≥50% official Kimi score is plausible by December 31, 2026. NOT FOUND: No confirmation that the official Scale/agi.safe.ai leaderboard itself displays or will display a tool-augmented Kimi score (its current listed Kimi K2.5 entry, 24.4%, is a no-tools figure); no evidence of a specific additional Kimi flagship (e.g., "K3.5" or "K4") confirmed for release before December 31, 2026, beyond unverified leak-stage rumors.

YES means: the official Humanity's Last Exam leaderboard at agi.safe.ai (or its confirmed successor source) shows some Moonshot Kimi model reaching an HLE accuracy of 50% or higher on or before December 31, 2026, 11:59 PM ET. - [2026-10-06] The authoritative Scale/CAIS leaderboard, mirrored via Klu's benchmark tracker, lists "Kimi K3 … Score: 22.2 overall" on Scale's Diamond split, the newest and largest Moonshot flagship, still less than half the 50% strike. https://klu.ai/benchmarks/humanitys-last-exam - [2026-09-30] A separate Scale snapshot confirms the same figure in ranked context: Kimi K3 sits at "22.2%" in 10th place among just 12 published rows, trailing even Grok 4.7's 23.4% and far behind leader GPT-6 Astra's 60.6%. https://benchlm.ai/benchmarks/scale-hle-diamond - [2026-10-04] The HLE-Diamond refresh (released Sept 22, 2026) explicitly lists "kimi-k3 22.2%" under its no-tools comparison table, showing the newer, harder Diamond variant did not lift Moonshot's score at all versus the original HLE. https://postcutoff.com/e/2026-09-22-hle-diamond/ - [2026-10-10] A completely independent aggregator (benchlm's model-ledger page) still shows Kimi K3's only "source-displayable" HLE entry is the self-reported 56% figure sourced explicitly to "Provider exact — Moonshot AI: Kimi K3 launch blog," not to any independently verified or official-leaderboard row — the gap between vendor claim and official score persists structurally. https://benchlm.ai/models/kimi-k3 - [2026-09-07] Independent tracker Artificial Analysis, checked six weeks after K3's July 16 launch, placed Kimi K3's overall Intelligence Index well behind rivals (44, versus GLM-5.3's 45), with a separately noted raw HLE figure "around 11%" for a comparable open model class, underscoring how far third-party scoring runs lag Moonshot's own marketing numbers. https://www.morphllm.com/kimi-k3 - [2026-08-14] As of mid-August 2026, no Kimi K4 (or any larger, unreleased Moonshot flagship that might plausibly jump the HLE score) had been announced: "Moonshot AI has made no announcement of a model named Kimi K4," with kimi-k3 still the newest model in Moonshot's own API docs. https://kimik4ai.com/models/kimi-k4/ - [2026-09-29] The sole public rumor of a next-generation Moonshot model (per a 29 July 2026 Information report) is limited to GPU-procurement chatter, with explicitly "zero real specifications" — no parameter count, benchmark, or release date — meaning there is no credible pipeline event in Q4 2026 that would materially raise Kimi's official score before the Dec 31 deadline. https://aitoolsreview.co.uk/insights/kimi-k4-release-date - [78 days ago, i.e. ~2026-07-25] A Manifold Markets contract on "Kimi K4 released in 2026?" was trading at only 13% chance, directly reflecting forecaster skepticism that any new, more-capable Moonshot flagship (the only plausible route to a 50%+ official HLE score) will ship before year-end. https://manifold.markets/Gen/kimi-k4-released-in-2026 - [2026-10-10] On the broader, non-Scale llm-stats.com aggregator (104 evaluated models, last updated October 2026), Kimi K3 ranks only 12th among tracked models with no tools score anywhere near the leaders, consistent across every independently-operated leaderboard checked. https://llm-stats.com/benchmarks/humanity's-last-exam - [2026-10-06] Even the most favorable third-party breakdown (Klu) reiterates there is "no single HLE leaderboard in 2026" and that Scale's full-set board — the market's literal resolution source — is the one on which Kimi sits at 22.2%, not the self-reported 56% figure that circulates in vendor and press coverage. https://klu.ai/benchmarks/humanitys-last-exam BOTTOM LINE: Every independently or officially tallied figure for any Kimi model — across three separate leaderboard refreshes (original HLE, HLE-Rolling, and the newly launched HLE-Diamond) spanning July through October 2026 — sits at 22–24%, roughly half the 50% strike, while the only path to 50%+ (a materially stronger unreleased flagship such as "K4") remains unannounced and rated unlikely by independent forecasters; absent a verified leaderboard submission crediting Moonshot's self-reported 50-56% tool-assisted figures, the resolution source gives no support for a YES outcome. NOT FOUND: No evidence of Moonshot formally submitting K2.6/K3 vendor-reported tool-assisted scores (50-56%) for independent verification on the official agi.safe.ai/Scale leaderboard; no confirmed Kimi K4 (or any Moonshot model beyond K3) release or benchmark as of this report's date.